Fig. 1

Illustration of the MuAt deep neural network to predict the type of a tumour from its catalogue of somatic mutations. First, mutation data is one-hot encoded. MuAt integrates three data modalities: 3-bp sequence motif, genomic position and genomic annotations. Then, embedded mutation vectors are fed to the attention mechanism. Finally, mutation-level features are combined into tumour-level features, and tumour type is predicted. MuAt models can be interrogated by analysing (1) the attention matrix to recover informative mutations for tumours and tumour types, (2) tumour-level features for tumour subtype discovery and (3) prediction performance